Gemini Minecraft run: 605 deaths before AI slew the Ender Dragon

The agents never saw a single frame of the game or touched a keyboard. Instead, every agent steered its character using Mineflayer, a free, open-source library for building Minecraft bots. To do anything, an agent wrote a short JavaScript program, such as walking to a spot and mining nearby iron, then read back its position, health, inventory or the error it hit.

The setup was deliberately bare:

• a randomly generated survival world on Easy difficulty;
• cheat commands switched off;
• one goal, defeat the Ender Dragon, with no assigned roles, strategy or game-specific tools.

The researchers say that, as far as they know, no AI team has ever finished the whole game on its own using just the standard bot library, with no scouted maps or human-written game helpers.

The most human-looking part is how the team organized itself. Nobody told the agents who should do what, yet they settled into roles within the run.

• janesick made tools, found a Nether fortress and tracked down the stronghold that hides the portal to the End. To find it, the agent tossed a pair of Eyes of Ender from two separate places, followed each flight with code and calculated the point where the lines met. Its estimate landed 43 meters from the portal room, which was 1.3 km away.
• gyldenia handled food and supplies.
• auratian lit the Nether portal and later cracked the key puzzle of the boss fight.
• bulicame helped build shelters against Endermen and landed the final blow.

They warned each other about threats, kept shared notes of coordinates and lessons, and went back to recover a fallen teammate’s items. That’s the kind of teamwork human Minecraft players argue about over voice chat.

The End took nearly 14 hours, more than half the whole run. The agents kept dying, 69 times to the dragon alone, because something kept healing it.

About 21 hours in, auratian figured it out: ten crystals sitting on tall pillars were restoring the dragon’s health. It also found the team could reach them by stacking dirt blocks. Over the next four hours and 40 minutes, the four destroyed all ten. The dragon died 39 minutes later, with bulicame clinging on at 2.5 of 20 health points.

Any experienced player knows to take out the crystals first. The point is that the agents had to discover it under pressure, after a long string of failures, and then change strategy. That ability to notice why a plan isn’t working is what separates a useful AI assistant from one that keeps repeating the same mistake.

Running AI agents for a day isn’t free, and the researchers published the bill. Up to the win, the four players processed about 1.35 billion input tokens, the chunks of text a model reads.

• To beat the dragon: about $520 at Gemini 4 Argon’s introductory list price, or $1,040 at the standard price.
• The whole session: about $810, or $1,620 at the standard price. That includes a fifth agent, the Observer, which watched for cheating and found none.

That’s a lot for a game you can buy for far less and finish yourself. But as a measure of what a day of autonomous AI work costs, it’s a useful benchmark for any business weighing agents against human time.

The paper is honest about the caveats, and they matter.

• The agents had x-ray knowledge. Mineflayer tells them exactly which blocks and creatures are nearby, even behind solid rock, and a plugin plans their routes. A human player has to explore and guess.
• Easy mode. The run used the game’s easiest survival setting.
• Lots of errors. About 22% of the agents’ programs failed.
• One run. The result comes from a single attempt, so we don’t know how often it would succeed.
• Not the first dragon kill. Two community projects reported earlier wins, but they used game-specific tools or a fixed, easier setup.

For years, game-playing AI meant custom systems built for one game, like DeepMind’s AlphaGo for the board game Go. What’s new here is a general-purpose model, the same kind that answers questions and writes code, figuring out a sprawling game with only a basic toolkit.

You won’t be hiring Gemini to grind your Minecraft world. The real target is longer, messier work. Maxence Faldor, one of the researchers, said on X that the lab wants to point this same approach at robotics, scientific research and programming. And Google co-founder Sergey Brin said last week that Gemini now helps build the next Gemini.

Safety got attention too. The agents ran in a sealed-off world, each program could only control its own player, and the team could pause or stop the run from a live dashboard. They never needed to. As AI agents start booking travel and editing files on real computers, those same guardrails, limited permissions and a human with a stop button, are the ones to look for.